2.2.1.

Procedure

Only participants who were randomized to the passive control group of the RCT, who received no intervention, were included. Participants were assessed on 2 consecutive days. Motor tests were assessed twice on day 1, before and after a resting period of 30 min while keeping the fNIRS cap and optodes in place. After a complete removal of the fNIRS system, the cap was re-attached the next day and the third test was conducted around the same time of day as the first measurement the day before. Cap position across days was standardized (see Sec. 2.2.4). Motor and fNIRS assessments were conducted using a block design with each block consisting of seven trials of 20 sec of rest followed by 20 sec of movement. Prior pilot testing showed an 80% true positive rate when using this design. The start of each trial was marked in the fNIRS signal with a task-synchronized trigger. The 20-sec resting period, conducted in stance for the postural task and in sitting for the tapping task, served as a baseline for the following movement trial. The postural task blocks were conducted before the tapping task in a fixed order.

2.2.2.

Postural task

The postural task consisted of a non-immersive virtual reality weight-shifting task,³⁰ as described in detail elsewhere (see Figure S1 in the Supplementary Material).³¹ In short, participants were standing in front of a screen at approximately three meters distance. They were instructed to shift their weight mediolaterally $>80\%$ of their a-priori determined individual limits of stability, which activated a virtual water jet. With the water jet, they attempted to hit as many virtual wasps as possible, which appeared on the left and right side of the screen alternately. Participants’ center of mass (CoM) was captured within Nexus software (Vicon, Oxford Metrics, United Kingdom) by recording reflective markers placed bilaterally on the acromia, posterior superior iliac spines, lateral epicondyles, and lateral malleoli. Calculation of the CoM was done online within the D-Flow software (Motek Medical BV, Amsterdam, The Netherlands; version 3.28), based on the formulation by Winter (2009).³² During rest, subjects were asked to stand and watch the screen, displaying a video recording of the wasp game.

2.2.3.

Finger tapping task

The finger-tapping task, as part of the cloud-UPDRS application (version 1.3.0),³³ was performed with the right dominant hand on a smartphone. It consisted of two visual targets with a diameter of 0.6 inch, positioned at a 2-inch distance from each other (see Figure S2 in the Supplementary Material). Participants were seated on a chair in front of the smartphone placed on a table. They were instructed to alternately tap the left and right targets with their right index finger while holding the smartphone still with their left hand. To prevent any learning effect, tapping frequency was set at 180 beats per minute imposed by a metronome beat. The researcher verbally indicated the start and termination of the task. Participants were instructed to sit as still as possible and place their right hand flat on the table during the 20-sec resting periods in between tapping trials.

2.2.4.

fNIRS assessment

Brain hemodynamics were recorded following recent consensus guidelines.⁵ A continuous wave, single-phase fNIRS system (NIRSport2, NIRx, Berlin, Germany), using light emitting diodes (LEDs) with wavelengths of 760 and 850 nm at a sampling frequency of 7.81 Hz, recorded brain hemodynamics within the Aurora software (version 2020.7). First, participants’ head circumference was measured and matched to the closest corresponding cap size, varying from 54 to 60 cm. The Cz anatomical landmark was then determined using the inion, nasion, and pre-auricular points after which the lightweight cap was carefully placed on the head. To ensure similar cap placement on day two, the Cz, CP2, and FC1 anatomical landmarks were marked. Participants were asked not to wash their hair in between measurement days.

A total of 32 optodes (16 sources, 16 detectors) were used to cover predefined ROIs, including the prefrontal cortex (PFC; Brodmann areas 9, 10, and 46), frontal eye fields (FEF; Brodmann area 8), supplementary motor area (SMA; Brodmann area 6, medial), premotor cortex (PMC; Brodmann area 6, lateral), and somatosensory cortex (SSC; Brodmann areas 1, 3, 5 and 7). The 10-10 layout in the fNIRS Optodes’ Location Decider (fOLD) toolbox³⁴ was used to specify the optode locations (see Table S1 in the Supplementary Material) with a source-detector separation of $\sim 3\mathrm{cm}$. Sixteen short-separation channels with a source-detector separation of 8 mm were included, one over each source, to correct for physiological noise in the fNIRS signal.³⁵ Furthermore, two accelerometers were placed at the back of the cap to correct for movement artefacts related to head movements. At the onset of testing, signal quality was visually checked and improved by moving the hair aside from underneath the optodes, if needed. An additional opaque cap was then placed over the fNIRS set-up to protect external light from interfering with the hemodynamic measurements. For the tapping task, the motor channel in between optode C3-C1 of the contralateral (left) hemisphere was chosen as the ROI. As the fOLD toolbox does not provide specific information on anatomical brain representations, the C3-C1 hand motor channel was based on the EEG topography³⁶ in accordance with previous data on the homunculus hand position.³⁷

2.3.1.

Behavioral data

Weight-shifting data were exported by D-Flow and analyzed within MATLAB 2018b (MathWorks, Natick, Massachusetts, United States). Similar to our previous analysis,³¹ CoM data were first low-pass filtered with a fourth-order Butterworth filter (cut-off = 6 Hz). Weight-shifting was then determined as the movement from the 80% stability limits threshold on the right to the 80% stability limits threshold on the left and vice versa. Outcome measures included weight-shifting speed and accuracy (CoM error) averaged over the seven trials.

The cloud-UPDRS smartphone data were analyzed with Microsoft Excel (version 2016). Outcome measures included the average number of taps per trial and the accuracy of target taps in pixels (target error) and calculated as:

2.3.2.

fNIRS data

Brain hemodynamic data were analyzed with the open access NIRS toolbox ( https://github.com/huppertt/nirs-toolbox)³⁸ implemented in MATLAB 2018b (MathWorks, Natick, Massachusetts, United States). Raw intensity signals were checked visually, resampled to 5 Hz, and converted into optical density using the Beer-Lambert law and a partial path length correction factor of 0.1, thereby correcting for light scatter caused by the brain tissue that the near-infrared light was travelling though, as stated in previous research.^39,40 A general linear model, including short separation channels and accelerometers, was used to estimate the task hemodynamic response, thereby correcting for signal variations due to physiological noise and movement artefacts.^20,35 The autoregressive iteratively reweighted least squares method was implemented to correct for motion and auto correlated noise.⁴¹ This method, including short separation channel regression, outperformed other filtering methods³⁵ and was shown to improve data reproducibility.³⁵ Accelerometers identified and corrected for changes in the fNIRS signal that corresponded with the accelerometer signal over a time-window of 15 s.⁴²

Channels were averaged within the predefined ROIs. A channel was included if the spatial specificity, as determined with the fOLD toolbox,³⁴ was at least 50% within a ROI⁴³ (see Fig. 1). As the SMA and PMC ROIs were grouped within the fOLD output, the medial channels were defined as SMA and the lateral channels as PMC.⁴⁴ The primary motor cortex (M1) was not included as ROI in this analysis, because spatial specificity did not exceed 50%. Midline-traversing channels were excluded, as cerebrospinal fluid running through the superior sagittal sinus could have interfered with the measurement of the underlying brain hemodynamics.⁴⁵ In addition, the tapping task-specific hand motor channel C3-C1 was excluded from the ROIs specified for the postural task, as spatial specificity was lower than 50% (35% M1 and 35% PMC). Trial-averaged relative oxygenated hemoglobin (${\mathrm{HbO}}_2$ active trial - ${\mathrm{HbO}}_2$ rest trial ($\mu \mathrm{mol}/\mathrm{L}$)) was used as the primary outcome for each ROI. As secondary outcomes, HHb ($\mu \mathrm{mol}/\mathrm{L}$) and ${\mathrm{HbO}}_2$ ($\mu \mathrm{mol}/\mathrm{L}$) values for the left and right ROIs were investigated separately.

Fig. 1

fNIRS lay-out with ROIs according to the fOLD toolbox for the postural task with a specificity level $\ge 50\%$. The PFC is represented by the blue colored channels, the FEF by the green channels, the SMA by the orange channels, the PMC by the purple channels, and the SSC by the yellow channels. The channels situated at the midline were excluded (indicated by black crosses). The C3-C1 channel (black circle; brown channel) was included as the task-specific hand motor channel for the finger tapping task, consistent with the EEG topography. Nz, naison; Iz, inion; LPA, left point auricular; and RPA, right point auricular.

To a-posteriori check which channels were active during the postural or the finger-tapping task, signal changes in individual channels were assessed in a complementary analysis. A channel was classified as active when there was a significant average change during the task compared to rest in both ${\mathrm{HbO}}_2$ (positive change) and HHb levels (negative change) over the three time points, which was FDR-corrected for the number of channels.

3.3.1.

ROI reliability during the postural task without cap repositioning

In the total ROIs (left plus right channels), no differences in ${\mathrm{HbO}}_2$ were found across time points [PFC: $F_{(1.41)}=0.62$, $p=0.49$; FEFs: $F_{(2)}=0.25$, $p=0.78$; SMA: $F_{(2)}=0.73$, $p=0.49$; PMC: $F_{(1.20)}=0.39$, $p=0.58$; SSC: $F_{(1.35)}=0.83$, $p=0.40$; see Figs. 3(a)–3(e)]. Exploratory post-hoc tests also did not show differences between the three tests (all corrected $\mathrm{p}\text{-}\text{values}>0.17$). The ${\mathrm{HbO}}_2$ test-retest reliability, as captured by the ICC-values and CIs between test 1-2 was excellent for the PFC ($\mathrm{ICC}=0.87$, $\mathrm{CI}=[0.71,0.95]$), PMC ($\mathrm{ICC}=0.78$, $\mathrm{CI}=[0.53,0.91]$) and SSC ($\mathrm{ICC}=0.78$, $\mathrm{CI}=[0.51,0.91]$), and fair for the FEF ($\mathrm{ICC}=0.51$, $\mathrm{CI}=[0.10,0.77]$) and SMA ($\mathrm{ICC}=0.48$, $\mathrm{CI}=[0.05,0.76]$) (see Table 2). Table 2 further illustrates that the reliability for the ROIs per hemisphere showed lower values, but still within an acceptable range (PFC: $\mathrm{ICC}\ge 0.76$; FEF: $\mathrm{ICC}\ge 0.49$; SMA: $\mathrm{ICC}\ge 0.40$; PMC: $\mathrm{ICC}\ge 0.74$; SSC: $\mathrm{ICC}\ge 0.60$). Bland–Altman plots are shown in Fig. 4. Points fell largely within the limits of agreement, were equally distributed around zero, and showed no bias, corroborating the findings above. The pattern, which was overall similar for the different ROIs, did show that one participant with extreme activation (PFC) or deactivation (PMC, SMA) values also had the least consistency between time points. As for the SEMs, values ranged between 1.89 and $4.09\mu \mathrm{mol}/\mathrm{L}$ for the total ROIs and between 2.51 and $4.91\mu \mathrm{mol}/\mathrm{L}$ for the ROIs per hemisphere.

Fig. 3

(a)–(e) Relative ${\mathrm{HbO}}_2$ levels (active trial - rest trial) for total (left plus right channels) ROIs at test 1, 2, and 3 for the postural task. Note that the scale on the $x$-axis differs between ROIs as to visualize the slightest differences. (f) Relative ${\mathrm{HbO}}_2$ levels for all total ROIs averaged over the three test moments. ROI, region of interest; PFC, prefrontal cortex; FEF, frontal eye fields; PMC, premotor cortex; SMA, supplementary motor area; and SSC, somatosensory cortex.

Table 2

Test-retest reliability based on single ICCs for relative HbO2 and HHb levels during the postural task.

Fig. 4

Bland–Altman plots for total (left plus right channels) ROIs during the postural task, visualizing the average ${\mathrm{HbO}}_2$ levels ($x$-axis) and the difference in ${\mathrm{HbO}}_2$ levels between test 1 and test 2 ($y$-axis). Dotted lines represent the 95% limits of agreement.

Overall, HHb test-retest reliability was lower compared to ${\mathrm{HbO}}_2$ ($\mathrm{ICC}-0.15$ on average for total ROIs), with some exceptions (see Table 2 and Fig. 2). Averaged ICCs, representing the test-retest reliability of the averaged test outcomes⁵⁰ showed higher values [see Table S2 in the Supplementary Material (${\mathrm{HbO}}_2$) and Table S3 in the Supplementary Material (HHb)]. Additionally, individual channel-based analysis revealed that 13 out of 32 channels were activated during the postural weight-shifting task [see Figure S4(a) in the Supplementary Material], including those in the SSC, SMA, and PMC.

3.3.2.

ROI reliability during the postural task with cap repositioning

Table 2 also illustrates that the repositioning of the cap between test 1 and 3 led to less stable test-retest values than between test 1 and 2. Reliability values were lower for the PFC ($\mathrm{ICC}=0.50$, $\mathrm{CI}=[0.08,0.77]$) and SSC ($\mathrm{ICC}=0.51$, $\mathrm{CI}=[0.09,0.78]$), but remained fair. Reliability became poor for the SMA ($\mathrm{ICC}=0.01$, $\mathrm{CI}=[0.00,0.45]$) and the PMC ($\mathrm{ICC}=0.00$, $\mathrm{CI}=[0.00,0.14]$). However, reliability was good for the FEF ($\mathrm{ICC}=0.64$, $\mathrm{CI}=[0.28,0.84]$). Table 2 also shows that reliability for the ROIs per hemisphere was similar for the motor areas (SMA: $\mathrm{ICC}\le 0.22$; PMC: $\mathrm{ICC}\le 0.01$), and lower for the FEF ($\mathrm{ICC}\le 0.48$). Interestingly, ICC-values were higher for the left PFC ($\mathrm{ICC}=0.66$, [0.32, 0.85]) and right SSC ($\mathrm{ICC}=0.65$, [0.31, 0.85]), but lower for the right PFC ($\mathrm{ICC}=0.00$, $\mathrm{CI}=[0.00,0.39]$) and left SSC ($\mathrm{ICC}=0.06$, [0.00, 0.49]). Bland–Altman plots of test 1 and 3 largely confirm the above described results (see Fig. 5). Limits of agreement increased for total ROIs compared to the test 1-2 comparison, especially for the PMC and SMA, but not for the FEF. The same participant with extreme (de)activation as previously mentioned, also showed the least consistency between tests 1 and 3 (PFC, PMC, SMA). SEM scores were also larger between test 1 and 3 (range: $2.47-5.73\mu \mathrm{mol}/\mathrm{L}$) than between test 1 and 2 (range: $2.51-4.09\mu \mathrm{mol}/\mathrm{L}$), with the FEF as the only exception (change in SEM: $-0.27\mu \mathrm{mol}/\mathrm{L}$). SEMs ranged between 3.18 and $7.30\mu \mathrm{mol}/\mathrm{L}$ for the ROIs per hemisphere.

Fig. 5

Bland–Altman plots for total (left plus right channels) ROIs during the postural task, visualizing the average ${\mathrm{HbO}}_2$ levels ($x$-axis) and the difference in ${\mathrm{HbO}}_2$ levels between test 1 and test 3 ($y$-axis). Dotted lines represent the 95% limits of agreement.

The overall HHb test-retest reliability was comparable to ${\mathrm{HbO}}_2$ ($\mathrm{ICC}-0.03$ on average for total ROIs), though less stable between test 1 and 3 than between test 1 and 2 (see Table 2 and Fig. 2). Averaged ICCs showed higher values for both ${\mathrm{HbO}}_2$ (see Table S2 in the Supplementary Material) and HHb (see Table S3 in the Supplementary Material).

3.3.3.

Task-specific fNIRS reliability during finger tapping

For the task-specific motor channel representing the right hand (C3-C1), no differences in relative ${\mathrm{HbO}}_2$ levels were found across time points during the tapping task [$F_{(1.53)}=1.75$, $p=0.20$; see Fig. 6(a)]. Exploratory post-hoc tests also did not show differences between test moments (all corrected $p\text{-}\text{values}>0.46$), even though a trend towards improved tapping accuracy was seen from test 1 to test 2. Interestingly, however, participants who showed better accuracy (lower tapping error) were as stable in their fNIRS outcomes, as assessed with Pearson’s correlations (${\mathrm{HbO}}_2$: $r=0.30$, $p=0.20$; HHb: $r=-0.33$, $p=0.15$; see Figure S5 in the Supplementary Material). The ICC-values were good between tests 1-2 ($\mathrm{ICC}=0.66$, $\mathrm{CI}=[0.31,0.84]$), but became poor between tests 1-3 ($\mathrm{ICC}=0.38$, $\mathrm{CI}=[0.00,0.70]$) (see Table 3). Bland–Altman plots largely confirm these results [see Figs. 6(b) and 6(c)]. Points fell between the limits of agreement, which covered a wider range and were more spread out between test 1 and 3 than between test 1 and 2. In agreement, the SEM score was $2.00\mu \mathrm{mol}/\mathrm{L}$ between test 1 and 2, and $2.93\mu \mathrm{mol}/\mathrm{L}$ between test 1 and 3.

Fig. 6

(a) Relative ${\mathrm{HbO}}_2$ levels (active trial - rest trial) at test 1, 2, and 3 for the hand motor channel C3-C1 during the finger tapping task; (b) Bland–Altman plot for channel C3-C1, visualizing the average ${\mathrm{HbO}}_2$ levels and the difference between ${\mathrm{HbO}}_2$ levels at test 1 and 2, and (c) test 2 and 3.

Table 3

Test-retest reliability based on single ICCs for relative HbO2 and HHb levels during the finger tapping task.

HHb test-retest reliability, as well as the ${\mathrm{HbO}}_2$ and HHb averaged ICCs (see Table S4 in the Supplementary Material), were better than the single ICCs for ${\mathrm{HbO}}_2$, ranging from fair to excellent. Additionally, individual channel-based analysis revealed that five channels of the left hemisphere were active during the right finger-tapping task [see Fig. S4(b) in the Supplementary Material], including the C3-C1 channel, which confirmed the anatomical location of the right-hand motor region.

4.2.1.

Reliability during postural control without cap removal

Test-retest analysis of five predefined and commonly applied ROIs during the postural task revealed an overall fair to excellent reliability in ${\mathrm{HbO}}_2$ levels when repeating two tests on the same day. This suggests that, when not removing the fNIRS cap and optodes, the fNIRS signal is fairly reliable over time for the FEF and SMA ($\mathrm{ICC}\ge 0.48$) and highly reliable for the PFC, PMC, and SSC ($\mathrm{ICC}\ge 0.78$). Similar results were found for the left and right hemispheres separately. To date, only one other study specifically investigated fNIRS test-retest reliability when testing twice on the same day (study 4 in Table 4),²⁷ although with cap removal. These authors studied only the PFC in young adults, resulting in a good reliability during both straight walking and turning ($\mathrm{ICC}\ge 0.67$). The higher reliability found in the current study for the postural task may be explained by not removing the cap, the lack of activity in this area, the shorter duration of signal extraction (20 sec versus 2 min) and higher number of trials (7 versus 1) in which the task was performed.⁵ Pilot testing prior to data collection also showed a 13% true positive rate increase by adding two trials (67% at five trials and 80% at seven trials). Additionally, other task-related aspects, such as heel strikes during gait, may induce movement artifacts, which were absent during the present postural task. As Stuart et al.²⁷ did not report on HHb, the overall lower reliability compared to ${\mathrm{HbO}}_2$ found in this study cannot directly be compared.

4.2.2.

Reliability during postural control with cap removal

An important finding of the present study is that when measuring fNIRS ${\mathrm{HbO}}_2$ levels on different days after cap removal and repositioning, reliability deteriorated. Interestingly, non-motor cortical areas seemed less affected by cap removal, indicated by fair to good reliability in the PFC, FEF, and SSC ($\mathrm{ICC}\ge 0.50$), which is comparable to the reliability found during walking²⁶ and handgrip exercises²³ (study 3 and 7 in Table 4). On the other hand, poor reliability was revealed in the motor areas ($\mathrm{ICC}\le 0.01$), in agreement with a reliability study during passive hand movements²¹ (study 2 in Table 4), possibly driven by the lack of active movement. The finding that the non-motor areas proved to be more stable across consecutive days could be due the fact that this virtual reality-based postural task required strong non-motor involvement. The relatively poor reliability found for the motor areas during the postural task ($\mathrm{ICC}\le 0.01$) is in contrast with previous research on hand grasping and finger tapping, which showed fair to excellent reliability after cap removal ($\mathrm{ICC}\ge 0.50$) (study 1, 5, and 6 in Table 4).^20,22,24 Besides cap repositioning, this finding may alternatively be explained by the fact that the postural task required whole body movements, resulting in wider and more variable cortical recruitment. The single channel-based analysis underscored this statement, as no clear activation pattern was found for the PMC and SMA. In addition, fNIRS signals pertaining to the lower limbs are more prone to systematic artifacts²⁹ and likely more difficult to capture as the motor representation of the lower extremities is located deeply within the interhemispheric fissure,³⁷ which could result in more inconsistent fNIRS measurements. Note that, similar to the present findings, HHb reliability after cap removal was overall comparable to ${\mathrm{HbO}}_2$ findings in healthy adults (see Table 4).

4.2.3.

Task-specific fNIRS reliability during finger tapping

For finger tapping, reliability of the C3-C1 hand motor channel was good when assessed on the same day ($\mathrm{ICC}=0.66$). This supports the notion that fNIRS is a sensitive neuroimaging tool for capturing ${\mathrm{HbO}}_2$ changes during a specific motor tasks requiring localized cortical activation.²⁹ It should be noted that there was a trend towards better behavioral tapping accuracy from test 1 to test 2 ($p=0.07$). However, when exploring this result, we did not find differences in the ${\mathrm{HbO}}_2$ levels or that the participants with improved accuracies had better reliability. Similar to the motor cortical ROIs of the postural task, reliability during tapping became poor when assessed on two consecutive days ($\mathrm{ICC}=0.38$), though the decline was less steep ($-0.47$ (SMA) and $-0.78$ (PMC) for the postural task and $-0.28$ (C3-C1) for the tapping task). This highlights that using fNIRS on 2 consecutive days results in suboptimal reproducibility, despite efforts to standardize cap replacement. In contrast, previous studies investigating task-specific motor areas showed fair to excellent ICC-values ($\ge 0.51$) during digit manipulation²² and hand grasping²⁰ when assessed on different days (study 1 and 5 in Table 4). These results, however, were found in young adults, who might show higher signal to noise ratios and therefore higher reliability.²⁸ It could also be that the localization method used in these studies, TMS localization²⁰ and registration of 3D-coordinates,²² provided a more accurate foundation for cap replacement after removal.

4.2.4.

Suggestions for improving fNIRS reliability

First, other methods of cap standardization could be explored, such as using a 3D-digitizer, a neuronavigation system⁵¹ or structural MRI scans for co-registration.⁵² Interestingly, we found that averaged ICCs were generally higher than single ICCs.^25,46 Especially for study protocols that involve cap removal, multiple fNIRS blocks within one testing session, and the use of averaged values of ${\mathrm{HbO}}_2$ and HHb levels for between-test comparisons may be required. Second, future studies should investigate whether shortening of trial duration while increasing the number of trials within a block design, could lead to enhanced stability of fNIRS outcomes, as pilot testing revealed increased positive rates using this procedure. Third, behavioral tasks should be standardized as much as possible and any learning effects counteracted, as was done in the present study. Measuring on the same time of day and including large sample sizes is also recommended, especially when using a complex (motor) task. Finally, person-specific variables, such as vigilance, motivation, and effort, could additionally be assessed, as they can possibly affect reliability.